An Algorithm Based on Deep Learning for Predicting In-Hospital Cardiac Arrest | JoVE Visualize

Area of Science:

Medical Informatics
Artificial Intelligence in Healthcare
Clinical Decision Support Systems

Background:

In-hospital cardiac arrest poses a significant threat to patient safety and public health.
Traditional track-and-trigger systems for early cardiac arrest detection have limitations, including low sensitivity and high false-alarm rates.
There is a need for more effective early warning systems to predict and prevent cardiac arrest.

Purpose of the Study:

To develop and evaluate a deep learning-based early warning system for predicting in-hospital cardiac arrest.
To compare the performance of the deep learning system against existing track-and-trigger systems and machine learning algorithms.
To assess the system's ability to maintain high sensitivity while reducing false alarms.

Main Methods:

A retrospective cohort study involving 52,131 patients admitted to two hospitals between June 2010 and July 2017.
A recurrent neural network (deep learning model) was trained on data from June 2010 to January 2017 and tested on data from February to July 2017.
Performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC), comparing the deep learning system to modified early warning score, random forest, and logistic regression.

Main Results:

The deep learning-based early warning system achieved a significantly higher AUROC (0.850) and AUPRC (0.044) compared to modified early warning score (AUROC: 0.603; AUPRC: 0.003), random forest (AUROC: 0.780; AUPRC: 0.014), and logistic regression (AUROC: 0.613; AUPRC: 0.007).
At equivalent sensitivity levels, the deep learning system reduced the number of alarms by 82.2% compared to the modified early warning score, 13.5% compared to random forest, and 42.1% compared to logistic regression.
The deep learning system demonstrated high sensitivity and a low false-alarm rate in detecting patients with cardiac arrest.

Conclusions:

A deep learning-based algorithm offers superior performance for early detection of in-hospital cardiac arrest.
The developed system shows potential to enhance patient safety by providing more accurate and timely warnings.
This approach significantly improves upon existing methods, offering a promising advancement in clinical decision support for critical events.